|
| 1 | +; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5 |
1 | 2 | ; RUN: opt -S -passes=loop-fusion -loop-fusion-peel-max-count=3 < %s | FileCheck %s
|
2 | 3 |
|
3 | 4 | ; Tests if we are able to fuse two guarded loops which have constant but
|
|
6 | 7 |
|
7 | 8 | @B = common global [1024 x i32] zeroinitializer, align 16
|
8 | 9 |
|
9 |
| -; CHECK-LABEL: void @main(ptr noalias %A) |
10 |
| -; CHECK-NEXT: entry: |
11 |
| -; CHECK: br i1 %cmp4, label %for.first.entry, label %for.end |
12 |
| -; CHECK: for.first.entry |
13 |
| -; CHECK-NEXT: br label %for.first.peel.begin |
14 |
| -; CHECK: for.first.peel.begin: |
15 |
| -; CHECK-NEXT: br label %for.first.peel |
16 |
| -; CHECK: for.first.peel: |
17 |
| -; CHECK: br label %for.first.peel.next |
18 |
| -; CHECK: for.first.peel.next: |
19 |
| -; CHECK-NEXT: br label %for.first.peel2 |
20 |
| -; CHECK: for.first.peel2: |
21 |
| -; CHECK: br label %for.first.peel.next1 |
22 |
| -; CHECK: for.first.peel.next1: |
23 |
| -; CHECK-NEXT: br label %for.first.peel.next11 |
24 |
| -; CHECK: for.first.peel.next11: |
25 |
| -; CHECK-NEXT: br label %for.first.entry.peel.newph |
26 |
| -; CHECK: for.first.entry.peel.newph: |
27 |
| -; CHECK: br label %for.first |
28 |
| -; CHECK: for.first: |
29 |
| -; CHECK: br i1 %cmp3, label %for.first, label %for.second.exit |
30 |
| -; CHECK: for.second.exit: |
31 |
| -; CHECK: br label %for.end |
32 |
| -; CHECK: for.end: |
33 |
| -; CHECK-NEXT: ret void |
34 | 10 |
|
35 | 11 | define void @main(ptr noalias %A) {
|
| 12 | +; CHECK-LABEL: define void @main( |
| 13 | +; CHECK-SAME: ptr noalias [[A:%.*]]) { |
| 14 | +; CHECK-NEXT: [[ENTRY:.*:]] |
| 15 | +; CHECK-NEXT: [[CMP4:%.*]] = icmp slt i64 0, 45 |
| 16 | +; CHECK-NEXT: [[CMP31:%.*]] = icmp slt i64 2, 45 |
| 17 | +; CHECK-NEXT: br i1 [[CMP4]], label %[[FOR_FIRST_ENTRY:.*]], label %[[FOR_END:.*]] |
| 18 | +; CHECK: [[FOR_FIRST_ENTRY]]: |
| 19 | +; CHECK-NEXT: br label %[[FOR_FIRST_PEEL_BEGIN:.*]] |
| 20 | +; CHECK: [[FOR_FIRST_PEEL_BEGIN]]: |
| 21 | +; CHECK-NEXT: br label %[[FOR_FIRST_PEEL:.*]] |
| 22 | +; CHECK: [[FOR_FIRST_PEEL]]: |
| 23 | +; CHECK-NEXT: [[SUB_PEEL:%.*]] = sub nsw i64 0, 3 |
| 24 | +; CHECK-NEXT: [[ADD_PEEL:%.*]] = add nsw i64 0, 3 |
| 25 | +; CHECK-NEXT: [[MUL_PEEL:%.*]] = mul nsw i64 [[SUB_PEEL]], [[ADD_PEEL]] |
| 26 | +; CHECK-NEXT: [[REM_PEEL:%.*]] = srem i64 [[MUL_PEEL]], 0 |
| 27 | +; CHECK-NEXT: [[CONV_PEEL:%.*]] = trunc i64 [[REM_PEEL]] to i32 |
| 28 | +; CHECK-NEXT: [[ARRAYIDX_PEEL:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 0 |
| 29 | +; CHECK-NEXT: store i32 [[CONV_PEEL]], ptr [[ARRAYIDX_PEEL]], align 4 |
| 30 | +; CHECK-NEXT: [[INC_PEEL:%.*]] = add nsw i64 0, 1 |
| 31 | +; CHECK-NEXT: [[CMP_PEEL:%.*]] = icmp slt i64 [[INC_PEEL]], 45 |
| 32 | +; CHECK-NEXT: br label %[[FOR_FIRST_PEEL_NEXT:.*]] |
| 33 | +; CHECK: [[FOR_FIRST_PEEL_NEXT]]: |
| 34 | +; CHECK-NEXT: br label %[[FOR_FIRST_PEEL2:.*]] |
| 35 | +; CHECK: [[FOR_FIRST_PEEL2]]: |
| 36 | +; CHECK-NEXT: [[SUB_PEEL3:%.*]] = sub nsw i64 [[INC_PEEL]], 3 |
| 37 | +; CHECK-NEXT: [[ADD_PEEL4:%.*]] = add nsw i64 [[INC_PEEL]], 3 |
| 38 | +; CHECK-NEXT: [[MUL_PEEL5:%.*]] = mul nsw i64 [[SUB_PEEL3]], [[ADD_PEEL4]] |
| 39 | +; CHECK-NEXT: [[REM_PEEL6:%.*]] = srem i64 [[MUL_PEEL5]], [[INC_PEEL]] |
| 40 | +; CHECK-NEXT: [[CONV_PEEL7:%.*]] = trunc i64 [[REM_PEEL6]] to i32 |
| 41 | +; CHECK-NEXT: [[ARRAYIDX_PEEL8:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[INC_PEEL]] |
| 42 | +; CHECK-NEXT: store i32 [[CONV_PEEL7]], ptr [[ARRAYIDX_PEEL8]], align 4 |
| 43 | +; CHECK-NEXT: [[INC_PEEL9:%.*]] = add nsw i64 [[INC_PEEL]], 1 |
| 44 | +; CHECK-NEXT: [[CMP_PEEL10:%.*]] = icmp slt i64 [[INC_PEEL9]], 45 |
| 45 | +; CHECK-NEXT: br label %[[FOR_FIRST_PEEL_NEXT1:.*]] |
| 46 | +; CHECK: [[FOR_FIRST_PEEL_NEXT1]]: |
| 47 | +; CHECK-NEXT: br label %[[FOR_FIRST_PEEL_NEXT11:.*]] |
| 48 | +; CHECK: [[FOR_FIRST_PEEL_NEXT11]]: |
| 49 | +; CHECK-NEXT: br label %[[FOR_FIRST_ENTRY_PEEL_NEWPH:.*]] |
| 50 | +; CHECK: [[FOR_FIRST_ENTRY_PEEL_NEWPH]]: |
| 51 | +; CHECK-NEXT: br label %[[FOR_FIRST:.*]] |
| 52 | +; CHECK: [[FOR_FIRST]]: |
| 53 | +; CHECK-NEXT: [[I_05:%.*]] = phi i64 [ [[INC:%.*]], %[[FOR_FIRST]] ], [ [[INC_PEEL9]], %[[FOR_FIRST_ENTRY_PEEL_NEWPH]] ] |
| 54 | +; CHECK-NEXT: [[I1_02:%.*]] = phi i64 [ [[INC14:%.*]], %[[FOR_FIRST]] ], [ 2, %[[FOR_FIRST_ENTRY_PEEL_NEWPH]] ] |
| 55 | +; CHECK-NEXT: [[SUB:%.*]] = sub nsw i64 [[I_05]], 3 |
| 56 | +; CHECK-NEXT: [[ADD:%.*]] = add nsw i64 [[I_05]], 3 |
| 57 | +; CHECK-NEXT: [[MUL:%.*]] = mul nsw i64 [[SUB]], [[ADD]] |
| 58 | +; CHECK-NEXT: [[REM:%.*]] = srem i64 [[MUL]], [[I_05]] |
| 59 | +; CHECK-NEXT: [[CONV:%.*]] = trunc i64 [[REM]] to i32 |
| 60 | +; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[I_05]] |
| 61 | +; CHECK-NEXT: store i32 [[CONV]], ptr [[ARRAYIDX]], align 4 |
| 62 | +; CHECK-NEXT: [[INC]] = add nsw i64 [[I_05]], 1 |
| 63 | +; CHECK-NEXT: [[CMP:%.*]] = icmp slt i64 [[INC]], 45 |
| 64 | +; CHECK-NEXT: [[SUB7:%.*]] = sub nsw i64 [[I1_02]], 3 |
| 65 | +; CHECK-NEXT: [[ADD8:%.*]] = add nsw i64 [[I1_02]], 3 |
| 66 | +; CHECK-NEXT: [[MUL9:%.*]] = mul nsw i64 [[SUB7]], [[ADD8]] |
| 67 | +; CHECK-NEXT: [[REM10:%.*]] = srem i64 [[MUL9]], [[I1_02]] |
| 68 | +; CHECK-NEXT: [[CONV11:%.*]] = trunc i64 [[REM10]] to i32 |
| 69 | +; CHECK-NEXT: [[ARRAYIDX12:%.*]] = getelementptr inbounds [1024 x i32], ptr @B, i64 0, i64 [[I1_02]] |
| 70 | +; CHECK-NEXT: store i32 [[CONV11]], ptr [[ARRAYIDX12]], align 4 |
| 71 | +; CHECK-NEXT: [[INC14]] = add nsw i64 [[I1_02]], 1 |
| 72 | +; CHECK-NEXT: [[CMP3:%.*]] = icmp slt i64 [[INC14]], 45 |
| 73 | +; CHECK-NEXT: br i1 [[CMP3]], label %[[FOR_FIRST]], label %[[FOR_SECOND_EXIT:.*]] |
| 74 | +; CHECK: [[FOR_SECOND_EXIT]]: |
| 75 | +; CHECK-NEXT: br label %[[FOR_END]] |
| 76 | +; CHECK: [[FOR_END]]: |
| 77 | +; CHECK-NEXT: ret void |
| 78 | +; |
36 | 79 | entry:
|
37 | 80 | %cmp4 = icmp slt i64 0, 45
|
38 | 81 | br i1 %cmp4, label %for.first.entry, label %for.second.guard
|
|
0 commit comments